<p>Influential points can distort the statistical inference, leading to misleading conclusions. Hence, influential diagnosis is an important issue in data analysis, but is less studied for high-dimensional data. When there are multiple influential observations, dealing with masking and swamping effects is challenging, and existing methods often impose strong distributional assumptions on non-influential observations such as the normality assumption. Moreover, these methods are sensitive to these assumptions. In this paper, we propose a distribution-insensitive influential measure, based on the correlation between predictors and a transformation of the response, which relaxes the assumption on the underlying distribution of non-influential observations significantly. Particularly, no assumption is imposed on the response except that the response is continuous. Furthermore, a distribution-insensitive influential point (DIP) detection method is proposed, which is efficient in handling masking and swamping effects and robust to the distribution of non-influential observations. Theoretical properties of DIP are established. Simulation results and real data analysis support the theoretical findings.</p>